Okta suggests filtering tool lists via identity permissions to lower AI agent prompt overhead and improve security.
Key facts
- •Okta's method filters MCP tool lists based on identity entitlements rather than using group-level gateway controls.
- •Internal modeling showed tool-schema costs can fall by more than 90% depending on the permission scenario.
- •The approach maps MCP tools to OAuth scopes to determine which tools a specific user or agent can access.
- •Okta performs scope checks at two stages: when the agent prompt is assembled and at the moment of tool execution.
- •The company claims this design reduces the 'blast radius' for compromised identities by limiting available actions.
Okta has introduced an identity-scoped approach to the Model Context Protocol (MCP) to reduce AI agent token consumption. By filtering the list of tools available to an agent based on user and agent permissions, the company aims to minimize the 'tool tax'—the prompt overhead caused by including unnecessary tool schemas in every model call.
Reducing the tool tax
AI agents often receive a full list of available tools, including names, descriptions, and parameters, in every prompt turn. Okta argues that this creates unnecessary token costs for tools that are never used. By applying identity-based scoping, Okta filters these lists before they reach the model, ensuring agents only see tools they are authorized to use.
Internal modeling results
In internal modeling, Okta found that scoping permissions reduced the number of visible tools by more than 90% in certain scenarios. The company stated that tool-schema token costs fell by a similar proportion. These models were based on Okta product data and public documentation, rather than live customer deployments.
Security and operational impact
Beyond cost savings, Okta positions this as a security measure that applies least-privilege access at the tool level. By removing unauthorized tools from an agent's view, the company aims to reduce the potential impact of a compromised identity. The system performs scope checks both when assembling the agent prompt and at runtime before a tool call executes.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by Artificial Intelligence News.
